fix(#62): consolidate mid-conversation system messages for strict chat templates

[Doorman11991]

Fixes #62.

Qwen3 / Qwen3.5 chat templates under llama.cpp --jinja raise System message must be at the beginning. and llama.cpp returns HTTP 400 — but only when tools are present, since that's when it compiles the template into a tool-call grammar.

SmallCode injects role: system content mid-conversation in several places (clarifier, plan request, planner injection, path-validation warnings, skill activation, compaction summaries), so the outgoing messages array routinely had system entries at positions other than 0.

Fix

• New src/session/message_normalizer.js#consolidateSystemMessages() collapses every system-role message into a single leading system message (order preserved, identical blocks de-duplicated) and keeps only non-system turns after it.
• Applied in both request builders (bin/smallcode.js and bin/model_client.js chatCompletion) immediately before the request body is built, so it catches stray system messages regardless of which path injected them.

Verification

• Instrumented the outgoing request bodies against a live Qwen3 model: every tool-bearing request now carries exactly one system message at index 0.
• 9 new unit tests in test/message_normalizer.test.js (merge, ordering, dedup, multimodal passthrough, idempotency, edge cases).
• Full suite: 157 passing. npm run build passes.